TLDR: A new automated workflow allows satellites to detect dynamic phenomena like volcanic plumes in real-time using onboard AI, then autonomously plan and execute high-resolution follow-up measurements, significantly increasing the scientific utility of observations compared to traditional methods.
A groundbreaking automated workflow is set to transform how satellites observe and measure dynamic science phenomena, such as volcanic plumes, from space. This innovative system, detailed in a recent research paper, leverages advanced onboard computing, computer vision, and machine learning to enable real-time detection and precise follow-up measurements of fleeting events.
Traditionally, the direction of satellite instruments has been a time-consuming process, often planned days or weeks in advance by ground operations teams. This delay makes it challenging to capture rapidly evolving events effectively. The new approach addresses this by integrating a two-stage observation process directly on the satellite.
How the System Works
The workflow begins with a wide field of view (WFOV) instrument capturing initial imagery. This data is then immediately analyzed onboard the satellite using sophisticated computer vision techniques to identify dynamic events. Once an event, like a volcanic plume, is detected, a narrow field of view (NFOV) instrument is autonomously directed to obtain high-resolution, pinpoint measurements. This entire sequence is designed to operate within seconds, meeting the critical time constraints required for observing fast-changing phenomena.
The researchers specifically applied this workflow to the observation of volcanic plumes. Accurate measurements of these plumes are vital for understanding volcanic processes and for issuing timely warnings about aviation hazards. The system aims to precisely measure the volume and height of these plumes from space. Beyond volcanoes, this workflow is versatile and can be adapted for other remote sensing applications, including the study of extraterrestrial plumes, wildfires, and harmful algal blooms.
Perception and Planning
The paper delves into the technical aspects of the system’s perception and planning capabilities. For plume detection, various classification methods were explored, ranging from traditional machine learning algorithms like Naive Bayes and Random Forest to advanced convolutional neural networks (CNNs) such as UNET-Uavsar and UNET-Xception. These models process raw, four-channel satellite images to generate a binary pixel mask that identifies the plume.
To ensure accuracy, a denoising process is applied to the classification outputs. This involves using morphological operations to merge disjoint plume polygons and filter out small, potentially irrelevant shapes, resulting in a cleaner and more reliable plume mask.
Following detection and denoising, specialized trajectory planning algorithms guide the NFOV sensor. Four distinct algorithms were developed: ‘Trace Outline’ for observing the plume’s boundary, ‘Track Center’ for sampling along its major axis, ‘Diagonal Transect’ for angled cross-sections, and ‘Lawnmower Transect’ for perpendicular cross-sections. Each algorithm is designed to gather specific types of data, maximizing the scientific return from the observations.
Also Read:
- Streamlining Methane Detection for Satellites with Machine Learning
- Autonomous Rovers Get Smarter: Prioritizing Scientific Targets with Depth and Curiosity
Significant Results
Simulations of the workflow yielded impressive results. The most effective classifiers, particularly the UNET architectures, demonstrated a substantial increase in the utility of observations compared to traditional baseline methods. For example, the UNET-Uavsar model achieved a remarkable 87% of observed pixels falling within the plume, a significant improvement over the 12% achieved by the best baseline method. The study also reported an order of magnitude increase in the ratio of plume observed and intensity utility, and a two orders of magnitude increase in gradient utility.
Crucially, the entire workflow, from initial image capture to autonomous trajectory planning, was shown to execute efficiently, with typical runtimes under 13 seconds. This efficiency meets the stringent real-time requirements for dynamic targeting in space.
This research marks a major advancement in autonomous remote sensing, promising more precise and valuable measurements of dynamic phenomena. The technology holds significant potential for future space missions, including NASA’s upcoming Federated Autonomous Measurement (FAME) demonstration, and its applications are expected to expand to various other scientific domains. For a deeper dive into the technical details, the full research paper is available here.
